src/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecEw_VdwLJ_GeomP1P1_avx_256_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_256_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_256_single.h"
  34 #include "kernelutil_x86_avx_256_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomP1P1_VF_avx_256_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEw_VdwLJ_GeomP1P1_VF_avx_256_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrE,jnrF,jnrG,jnrH;
  62     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  63     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  64     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  65     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  66     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  67     real             rcutoff_scalar;
  68     real             *shiftvec,*fshift,*x,*f;
  69     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  70     real             scratch[4*DIM];
  71     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  72     real *           vdwioffsetptr0;
  73     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  74     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  75     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  77     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  78     real             *charge;
  79     int              nvdwtype;
  80     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  81     int              *vdwtype;
  82     real             *vdwparam;
  83     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
  84     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
  85     __m256i          ewitab;
  86     __m128i          ewitab_lo,ewitab_hi;
  87     __m256           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  88     __m256           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  89     real             *ewtab;
  90     __m256           dummy_mask,cutoff_mask;
  91     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
  92     __m256           one     = _mm256_set1_ps(1.0);
  93     __m256           two     = _mm256_set1_ps(2.0);
  94     x                = xx[0];
  95     f                = ff[0];
  96
  97     nri              = nlist->nri;
  98     iinr             = nlist->iinr;
  99     jindex           = nlist->jindex;
 100     jjnr             = nlist->jjnr;
 101     shiftidx         = nlist->shift;
 102     gid              = nlist->gid;
 103     shiftvec         = fr->shift_vec[0];
 104     fshift           = fr->fshift[0];
 105     facel            = _mm256_set1_ps(fr->epsfac);
 106     charge           = mdatoms->chargeA;
 107     nvdwtype         = fr->ntype;
 108     vdwparam         = fr->nbfp;
 109     vdwtype          = mdatoms->typeA;
 110
 111     sh_ewald         = _mm256_set1_ps(fr->ic->sh_ewald);
 112     beta             = _mm256_set1_ps(fr->ic->ewaldcoeff);
 113     beta2            = _mm256_mul_ps(beta,beta);
 114     beta3            = _mm256_mul_ps(beta,beta2);
 115
 116     ewtab            = fr->ic->tabq_coul_FDV0;
 117     ewtabscale       = _mm256_set1_ps(fr->ic->tabq_scale);
 118     ewtabhalfspace   = _mm256_set1_ps(0.5/fr->ic->tabq_scale);
 119
 120     /* Avoid stupid compiler warnings */
 121     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 122     j_coord_offsetA = 0;
 123     j_coord_offsetB = 0;
 124     j_coord_offsetC = 0;
 125     j_coord_offsetD = 0;
 126     j_coord_offsetE = 0;
 127     j_coord_offsetF = 0;
 128     j_coord_offsetG = 0;
 129     j_coord_offsetH = 0;
 130
 131     outeriter        = 0;
 132     inneriter        = 0;
 133
 134     for(iidx=0;iidx<4*DIM;iidx++)
 135     {
 136         scratch[iidx] = 0.0;
 137     }
 138
 139     /* Start outer loop over neighborlists */
 140     for(iidx=0; iidx<nri; iidx++)
 141     {
 142         /* Load shift vector for this list */
 143         i_shift_offset   = DIM*shiftidx[iidx];
 144
 145         /* Load limits for loop over neighbors */
 146         j_index_start    = jindex[iidx];
 147         j_index_end      = jindex[iidx+1];
 148
 149         /* Get outer coordinate index */
 150         inr              = iinr[iidx];
 151         i_coord_offset   = DIM*inr;
 152
 153         /* Load i particle coords and add shift vector */
 154         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 155
 156         fix0             = _mm256_setzero_ps();
 157         fiy0             = _mm256_setzero_ps();
 158         fiz0             = _mm256_setzero_ps();
 159
 160         /* Load parameters for i particles */
 161         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 162         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 163
 164         /* Reset potential sums */
 165         velecsum         = _mm256_setzero_ps();
 166         vvdwsum          = _mm256_setzero_ps();
 167
 168         /* Start inner kernel loop */
 169         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 170         {
 171
 172             /* Get j neighbor index, and coordinate index */
 173             jnrA             = jjnr[jidx];
 174             jnrB             = jjnr[jidx+1];
 175             jnrC             = jjnr[jidx+2];
 176             jnrD             = jjnr[jidx+3];
 177             jnrE             = jjnr[jidx+4];
 178             jnrF             = jjnr[jidx+5];
 179             jnrG             = jjnr[jidx+6];
 180             jnrH             = jjnr[jidx+7];
 181             j_coord_offsetA  = DIM*jnrA;
 182             j_coord_offsetB  = DIM*jnrB;
 183             j_coord_offsetC  = DIM*jnrC;
 184             j_coord_offsetD  = DIM*jnrD;
 185             j_coord_offsetE  = DIM*jnrE;
 186             j_coord_offsetF  = DIM*jnrF;
 187             j_coord_offsetG  = DIM*jnrG;
 188             j_coord_offsetH  = DIM*jnrH;
 189
 190             /* load j atom coordinates */
 191             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 192                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 193                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 194                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 195                                                  &jx0,&jy0,&jz0);
 196
 197             /* Calculate displacement vector */
 198             dx00             = _mm256_sub_ps(ix0,jx0);
 199             dy00             = _mm256_sub_ps(iy0,jy0);
 200             dz00             = _mm256_sub_ps(iz0,jz0);
 201
 202             /* Calculate squared distance and things based on it */
 203             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 204
 205             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 206
 207             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 208
 209             /* Load parameters for j particles */
 210             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 211                                                                  charge+jnrC+0,charge+jnrD+0,
 212                                                                  charge+jnrE+0,charge+jnrF+0,
 213                                                                  charge+jnrG+0,charge+jnrH+0);
 214             vdwjidx0A        = 2*vdwtype[jnrA+0];
 215             vdwjidx0B        = 2*vdwtype[jnrB+0];
 216             vdwjidx0C        = 2*vdwtype[jnrC+0];
 217             vdwjidx0D        = 2*vdwtype[jnrD+0];
 218             vdwjidx0E        = 2*vdwtype[jnrE+0];
 219             vdwjidx0F        = 2*vdwtype[jnrF+0];
 220             vdwjidx0G        = 2*vdwtype[jnrG+0];
 221             vdwjidx0H        = 2*vdwtype[jnrH+0];
 222
 223             /**************************
 224              * CALCULATE INTERACTIONS *
 225              **************************/
 226
 227             r00              = _mm256_mul_ps(rsq00,rinv00);
 228
 229             /* Compute parameters for interactions between i and j atoms */
 230             qq00             = _mm256_mul_ps(iq0,jq0);
 231             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 232                                             vdwioffsetptr0+vdwjidx0B,
 233                                             vdwioffsetptr0+vdwjidx0C,
 234                                             vdwioffsetptr0+vdwjidx0D,
 235                                             vdwioffsetptr0+vdwjidx0E,
 236                                             vdwioffsetptr0+vdwjidx0F,
 237                                             vdwioffsetptr0+vdwjidx0G,
 238                                             vdwioffsetptr0+vdwjidx0H,
 239                                             &c6_00,&c12_00);
 240
 241             /* EWALD ELECTROSTATICS */
 242
 243             /* Analytical PME correction */
 244             zeta2            = _mm256_mul_ps(beta2,rsq00);
 245             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 246             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 247             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 248             felec            = _mm256_mul_ps(qq00,felec);
 249             pmecorrV         = gmx_mm256_pmecorrV_ps(zeta2);
 250             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 251             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 252             velec            = _mm256_mul_ps(qq00,velec);
 253
 254             /* LENNARD-JONES DISPERSION/REPULSION */
 255
 256             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 257             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 258             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 259             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 260             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 261
 262             /* Update potential sum for this i atom from the interaction with this j atom. */
 263             velecsum         = _mm256_add_ps(velecsum,velec);
 264             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 265
 266             fscal            = _mm256_add_ps(felec,fvdw);
 267
 268             /* Calculate temporary vectorial force */
 269             tx               = _mm256_mul_ps(fscal,dx00);
 270             ty               = _mm256_mul_ps(fscal,dy00);
 271             tz               = _mm256_mul_ps(fscal,dz00);
 272
 273             /* Update vectorial force */
 274             fix0             = _mm256_add_ps(fix0,tx);
 275             fiy0             = _mm256_add_ps(fiy0,ty);
 276             fiz0             = _mm256_add_ps(fiz0,tz);
 277
 278             fjptrA             = f+j_coord_offsetA;
 279             fjptrB             = f+j_coord_offsetB;
 280             fjptrC             = f+j_coord_offsetC;
 281             fjptrD             = f+j_coord_offsetD;
 282             fjptrE             = f+j_coord_offsetE;
 283             fjptrF             = f+j_coord_offsetF;
 284             fjptrG             = f+j_coord_offsetG;
 285             fjptrH             = f+j_coord_offsetH;
 286             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 287
 288             /* Inner loop uses 96 flops */
 289         }
 290
 291         if(jidx<j_index_end)
 292         {
 293
 294             /* Get j neighbor index, and coordinate index */
 295             jnrlistA         = jjnr[jidx];
 296             jnrlistB         = jjnr[jidx+1];
 297             jnrlistC         = jjnr[jidx+2];
 298             jnrlistD         = jjnr[jidx+3];
 299             jnrlistE         = jjnr[jidx+4];
 300             jnrlistF         = jjnr[jidx+5];
 301             jnrlistG         = jjnr[jidx+6];
 302             jnrlistH         = jjnr[jidx+7];
 303             /* Sign of each element will be negative for non-real atoms.
 304              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 305              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 306              */
 307             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 308                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 309
 310             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 311             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 312             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 313             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 314             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 315             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 316             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 317             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 318             j_coord_offsetA  = DIM*jnrA;
 319             j_coord_offsetB  = DIM*jnrB;
 320             j_coord_offsetC  = DIM*jnrC;
 321             j_coord_offsetD  = DIM*jnrD;
 322             j_coord_offsetE  = DIM*jnrE;
 323             j_coord_offsetF  = DIM*jnrF;
 324             j_coord_offsetG  = DIM*jnrG;
 325             j_coord_offsetH  = DIM*jnrH;
 326
 327             /* load j atom coordinates */
 328             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 329                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 330                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 331                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 332                                                  &jx0,&jy0,&jz0);
 333
 334             /* Calculate displacement vector */
 335             dx00             = _mm256_sub_ps(ix0,jx0);
 336             dy00             = _mm256_sub_ps(iy0,jy0);
 337             dz00             = _mm256_sub_ps(iz0,jz0);
 338
 339             /* Calculate squared distance and things based on it */
 340             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 341
 342             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 343
 344             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 345
 346             /* Load parameters for j particles */
 347             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 348                                                                  charge+jnrC+0,charge+jnrD+0,
 349                                                                  charge+jnrE+0,charge+jnrF+0,
 350                                                                  charge+jnrG+0,charge+jnrH+0);
 351             vdwjidx0A        = 2*vdwtype[jnrA+0];
 352             vdwjidx0B        = 2*vdwtype[jnrB+0];
 353             vdwjidx0C        = 2*vdwtype[jnrC+0];
 354             vdwjidx0D        = 2*vdwtype[jnrD+0];
 355             vdwjidx0E        = 2*vdwtype[jnrE+0];
 356             vdwjidx0F        = 2*vdwtype[jnrF+0];
 357             vdwjidx0G        = 2*vdwtype[jnrG+0];
 358             vdwjidx0H        = 2*vdwtype[jnrH+0];
 359
 360             /**************************
 361              * CALCULATE INTERACTIONS *
 362              **************************/
 363
 364             r00              = _mm256_mul_ps(rsq00,rinv00);
 365             r00              = _mm256_andnot_ps(dummy_mask,r00);
 366
 367             /* Compute parameters for interactions between i and j atoms */
 368             qq00             = _mm256_mul_ps(iq0,jq0);
 369             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 370                                             vdwioffsetptr0+vdwjidx0B,
 371                                             vdwioffsetptr0+vdwjidx0C,
 372                                             vdwioffsetptr0+vdwjidx0D,
 373                                             vdwioffsetptr0+vdwjidx0E,
 374                                             vdwioffsetptr0+vdwjidx0F,
 375                                             vdwioffsetptr0+vdwjidx0G,
 376                                             vdwioffsetptr0+vdwjidx0H,
 377                                             &c6_00,&c12_00);
 378
 379             /* EWALD ELECTROSTATICS */
 380
 381             /* Analytical PME correction */
 382             zeta2            = _mm256_mul_ps(beta2,rsq00);
 383             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 384             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 385             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 386             felec            = _mm256_mul_ps(qq00,felec);
 387             pmecorrV         = gmx_mm256_pmecorrV_ps(zeta2);
 388             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 389             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 390             velec            = _mm256_mul_ps(qq00,velec);
 391
 392             /* LENNARD-JONES DISPERSION/REPULSION */
 393
 394             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 395             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 396             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 397             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 398             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 399
 400             /* Update potential sum for this i atom from the interaction with this j atom. */
 401             velec            = _mm256_andnot_ps(dummy_mask,velec);
 402             velecsum         = _mm256_add_ps(velecsum,velec);
 403             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 404             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 405
 406             fscal            = _mm256_add_ps(felec,fvdw);
 407
 408             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 409
 410             /* Calculate temporary vectorial force */
 411             tx               = _mm256_mul_ps(fscal,dx00);
 412             ty               = _mm256_mul_ps(fscal,dy00);
 413             tz               = _mm256_mul_ps(fscal,dz00);
 414
 415             /* Update vectorial force */
 416             fix0             = _mm256_add_ps(fix0,tx);
 417             fiy0             = _mm256_add_ps(fiy0,ty);
 418             fiz0             = _mm256_add_ps(fiz0,tz);
 419
 420             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 421             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 422             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 423             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 424             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 425             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 426             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 427             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 428             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 429
 430             /* Inner loop uses 97 flops */
 431         }
 432
 433         /* End of innermost loop */
 434
 435         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 436                                                  f+i_coord_offset,fshift+i_shift_offset);
 437
 438         ggid                        = gid[iidx];
 439         /* Update potential energies */
 440         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 441         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 442
 443         /* Increment number of inner iterations */
 444         inneriter                  += j_index_end - j_index_start;
 445
 446         /* Outer loop uses 9 flops */
 447     }
 448
 449     /* Increment number of outer iterations */
 450     outeriter        += nri;
 451
 452     /* Update outer/inner flops */
 453
 454     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*97);
 455 }
 456 /*
 457  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_256_single
 458  * Electrostatics interaction: Ewald
 459  * VdW interaction:            LennardJones
 460  * Geometry:                   Particle-Particle
 461  * Calculate force/pot:        Force
 462  */
 463 void
 464 nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_256_single
 465                     (t_nblist * gmx_restrict                nlist,
 466                      rvec * gmx_restrict                    xx,
 467                      rvec * gmx_restrict                    ff,
 468                      t_forcerec * gmx_restrict              fr,
 469                      t_mdatoms * gmx_restrict               mdatoms,
 470                      nb_kernel_data_t * gmx_restrict        kernel_data,
 471                      t_nrnb * gmx_restrict                  nrnb)
 472 {
 473     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 474      * just 0 for non-waters.
 475      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 476      * jnr indices corresponding to data put in the four positions in the SIMD register.
 477      */
 478     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 479     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 480     int              jnrA,jnrB,jnrC,jnrD;
 481     int              jnrE,jnrF,jnrG,jnrH;
 482     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 483     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 484     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 485     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 486     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 487     real             rcutoff_scalar;
 488     real             *shiftvec,*fshift,*x,*f;
 489     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 490     real             scratch[4*DIM];
 491     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 492     real *           vdwioffsetptr0;
 493     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 494     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 495     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 496     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 497     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 498     real             *charge;
 499     int              nvdwtype;
 500     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 501     int              *vdwtype;
 502     real             *vdwparam;
 503     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 504     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 505     __m256i          ewitab;
 506     __m128i          ewitab_lo,ewitab_hi;
 507     __m256           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 508     __m256           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 509     real             *ewtab;
 510     __m256           dummy_mask,cutoff_mask;
 511     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 512     __m256           one     = _mm256_set1_ps(1.0);
 513     __m256           two     = _mm256_set1_ps(2.0);
 514     x                = xx[0];
 515     f                = ff[0];
 516
 517     nri              = nlist->nri;
 518     iinr             = nlist->iinr;
 519     jindex           = nlist->jindex;
 520     jjnr             = nlist->jjnr;
 521     shiftidx         = nlist->shift;
 522     gid              = nlist->gid;
 523     shiftvec         = fr->shift_vec[0];
 524     fshift           = fr->fshift[0];
 525     facel            = _mm256_set1_ps(fr->epsfac);
 526     charge           = mdatoms->chargeA;
 527     nvdwtype         = fr->ntype;
 528     vdwparam         = fr->nbfp;
 529     vdwtype          = mdatoms->typeA;
 530
 531     sh_ewald         = _mm256_set1_ps(fr->ic->sh_ewald);
 532     beta             = _mm256_set1_ps(fr->ic->ewaldcoeff);
 533     beta2            = _mm256_mul_ps(beta,beta);
 534     beta3            = _mm256_mul_ps(beta,beta2);
 535
 536     ewtab            = fr->ic->tabq_coul_F;
 537     ewtabscale       = _mm256_set1_ps(fr->ic->tabq_scale);
 538     ewtabhalfspace   = _mm256_set1_ps(0.5/fr->ic->tabq_scale);
 539
 540     /* Avoid stupid compiler warnings */
 541     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 542     j_coord_offsetA = 0;
 543     j_coord_offsetB = 0;
 544     j_coord_offsetC = 0;
 545     j_coord_offsetD = 0;
 546     j_coord_offsetE = 0;
 547     j_coord_offsetF = 0;
 548     j_coord_offsetG = 0;
 549     j_coord_offsetH = 0;
 550
 551     outeriter        = 0;
 552     inneriter        = 0;
 553
 554     for(iidx=0;iidx<4*DIM;iidx++)
 555     {
 556         scratch[iidx] = 0.0;
 557     }
 558
 559     /* Start outer loop over neighborlists */
 560     for(iidx=0; iidx<nri; iidx++)
 561     {
 562         /* Load shift vector for this list */
 563         i_shift_offset   = DIM*shiftidx[iidx];
 564
 565         /* Load limits for loop over neighbors */
 566         j_index_start    = jindex[iidx];
 567         j_index_end      = jindex[iidx+1];
 568
 569         /* Get outer coordinate index */
 570         inr              = iinr[iidx];
 571         i_coord_offset   = DIM*inr;
 572
 573         /* Load i particle coords and add shift vector */
 574         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 575
 576         fix0             = _mm256_setzero_ps();
 577         fiy0             = _mm256_setzero_ps();
 578         fiz0             = _mm256_setzero_ps();
 579
 580         /* Load parameters for i particles */
 581         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 582         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 583
 584         /* Start inner kernel loop */
 585         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 586         {
 587
 588             /* Get j neighbor index, and coordinate index */
 589             jnrA             = jjnr[jidx];
 590             jnrB             = jjnr[jidx+1];
 591             jnrC             = jjnr[jidx+2];
 592             jnrD             = jjnr[jidx+3];
 593             jnrE             = jjnr[jidx+4];
 594             jnrF             = jjnr[jidx+5];
 595             jnrG             = jjnr[jidx+6];
 596             jnrH             = jjnr[jidx+7];
 597             j_coord_offsetA  = DIM*jnrA;
 598             j_coord_offsetB  = DIM*jnrB;
 599             j_coord_offsetC  = DIM*jnrC;
 600             j_coord_offsetD  = DIM*jnrD;
 601             j_coord_offsetE  = DIM*jnrE;
 602             j_coord_offsetF  = DIM*jnrF;
 603             j_coord_offsetG  = DIM*jnrG;
 604             j_coord_offsetH  = DIM*jnrH;
 605
 606             /* load j atom coordinates */
 607             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 608                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 609                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 610                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 611                                                  &jx0,&jy0,&jz0);
 612
 613             /* Calculate displacement vector */
 614             dx00             = _mm256_sub_ps(ix0,jx0);
 615             dy00             = _mm256_sub_ps(iy0,jy0);
 616             dz00             = _mm256_sub_ps(iz0,jz0);
 617
 618             /* Calculate squared distance and things based on it */
 619             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 620
 621             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 622
 623             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 624
 625             /* Load parameters for j particles */
 626             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 627                                                                  charge+jnrC+0,charge+jnrD+0,
 628                                                                  charge+jnrE+0,charge+jnrF+0,
 629                                                                  charge+jnrG+0,charge+jnrH+0);
 630             vdwjidx0A        = 2*vdwtype[jnrA+0];
 631             vdwjidx0B        = 2*vdwtype[jnrB+0];
 632             vdwjidx0C        = 2*vdwtype[jnrC+0];
 633             vdwjidx0D        = 2*vdwtype[jnrD+0];
 634             vdwjidx0E        = 2*vdwtype[jnrE+0];
 635             vdwjidx0F        = 2*vdwtype[jnrF+0];
 636             vdwjidx0G        = 2*vdwtype[jnrG+0];
 637             vdwjidx0H        = 2*vdwtype[jnrH+0];
 638
 639             /**************************
 640              * CALCULATE INTERACTIONS *
 641              **************************/
 642
 643             r00              = _mm256_mul_ps(rsq00,rinv00);
 644
 645             /* Compute parameters for interactions between i and j atoms */
 646             qq00             = _mm256_mul_ps(iq0,jq0);
 647             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 648                                             vdwioffsetptr0+vdwjidx0B,
 649                                             vdwioffsetptr0+vdwjidx0C,
 650                                             vdwioffsetptr0+vdwjidx0D,
 651                                             vdwioffsetptr0+vdwjidx0E,
 652                                             vdwioffsetptr0+vdwjidx0F,
 653                                             vdwioffsetptr0+vdwjidx0G,
 654                                             vdwioffsetptr0+vdwjidx0H,
 655                                             &c6_00,&c12_00);
 656
 657             /* EWALD ELECTROSTATICS */
 658
 659             /* Analytical PME correction */
 660             zeta2            = _mm256_mul_ps(beta2,rsq00);
 661             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 662             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 663             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 664             felec            = _mm256_mul_ps(qq00,felec);
 665
 666             /* LENNARD-JONES DISPERSION/REPULSION */
 667
 668             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 669             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 670
 671             fscal            = _mm256_add_ps(felec,fvdw);
 672
 673             /* Calculate temporary vectorial force */
 674             tx               = _mm256_mul_ps(fscal,dx00);
 675             ty               = _mm256_mul_ps(fscal,dy00);
 676             tz               = _mm256_mul_ps(fscal,dz00);
 677
 678             /* Update vectorial force */
 679             fix0             = _mm256_add_ps(fix0,tx);
 680             fiy0             = _mm256_add_ps(fiy0,ty);
 681             fiz0             = _mm256_add_ps(fiz0,tz);
 682
 683             fjptrA             = f+j_coord_offsetA;
 684             fjptrB             = f+j_coord_offsetB;
 685             fjptrC             = f+j_coord_offsetC;
 686             fjptrD             = f+j_coord_offsetD;
 687             fjptrE             = f+j_coord_offsetE;
 688             fjptrF             = f+j_coord_offsetF;
 689             fjptrG             = f+j_coord_offsetG;
 690             fjptrH             = f+j_coord_offsetH;
 691             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 692
 693             /* Inner loop uses 63 flops */
 694         }
 695
 696         if(jidx<j_index_end)
 697         {
 698
 699             /* Get j neighbor index, and coordinate index */
 700             jnrlistA         = jjnr[jidx];
 701             jnrlistB         = jjnr[jidx+1];
 702             jnrlistC         = jjnr[jidx+2];
 703             jnrlistD         = jjnr[jidx+3];
 704             jnrlistE         = jjnr[jidx+4];
 705             jnrlistF         = jjnr[jidx+5];
 706             jnrlistG         = jjnr[jidx+6];
 707             jnrlistH         = jjnr[jidx+7];
 708             /* Sign of each element will be negative for non-real atoms.
 709              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 710              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 711              */
 712             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 713                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 714
 715             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 716             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 717             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 718             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 719             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 720             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 721             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 722             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 723             j_coord_offsetA  = DIM*jnrA;
 724             j_coord_offsetB  = DIM*jnrB;
 725             j_coord_offsetC  = DIM*jnrC;
 726             j_coord_offsetD  = DIM*jnrD;
 727             j_coord_offsetE  = DIM*jnrE;
 728             j_coord_offsetF  = DIM*jnrF;
 729             j_coord_offsetG  = DIM*jnrG;
 730             j_coord_offsetH  = DIM*jnrH;
 731
 732             /* load j atom coordinates */
 733             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 734                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 735                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 736                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 737                                                  &jx0,&jy0,&jz0);
 738
 739             /* Calculate displacement vector */
 740             dx00             = _mm256_sub_ps(ix0,jx0);
 741             dy00             = _mm256_sub_ps(iy0,jy0);
 742             dz00             = _mm256_sub_ps(iz0,jz0);
 743
 744             /* Calculate squared distance and things based on it */
 745             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 746
 747             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 748
 749             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 750
 751             /* Load parameters for j particles */
 752             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 753                                                                  charge+jnrC+0,charge+jnrD+0,
 754                                                                  charge+jnrE+0,charge+jnrF+0,
 755                                                                  charge+jnrG+0,charge+jnrH+0);
 756             vdwjidx0A        = 2*vdwtype[jnrA+0];
 757             vdwjidx0B        = 2*vdwtype[jnrB+0];
 758             vdwjidx0C        = 2*vdwtype[jnrC+0];
 759             vdwjidx0D        = 2*vdwtype[jnrD+0];
 760             vdwjidx0E        = 2*vdwtype[jnrE+0];
 761             vdwjidx0F        = 2*vdwtype[jnrF+0];
 762             vdwjidx0G        = 2*vdwtype[jnrG+0];
 763             vdwjidx0H        = 2*vdwtype[jnrH+0];
 764
 765             /**************************
 766              * CALCULATE INTERACTIONS *
 767              **************************/
 768
 769             r00              = _mm256_mul_ps(rsq00,rinv00);
 770             r00              = _mm256_andnot_ps(dummy_mask,r00);
 771
 772             /* Compute parameters for interactions between i and j atoms */
 773             qq00             = _mm256_mul_ps(iq0,jq0);
 774             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 775                                             vdwioffsetptr0+vdwjidx0B,
 776                                             vdwioffsetptr0+vdwjidx0C,
 777                                             vdwioffsetptr0+vdwjidx0D,
 778                                             vdwioffsetptr0+vdwjidx0E,
 779                                             vdwioffsetptr0+vdwjidx0F,
 780                                             vdwioffsetptr0+vdwjidx0G,
 781                                             vdwioffsetptr0+vdwjidx0H,
 782                                             &c6_00,&c12_00);
 783
 784             /* EWALD ELECTROSTATICS */
 785
 786             /* Analytical PME correction */
 787             zeta2            = _mm256_mul_ps(beta2,rsq00);
 788             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 789             pmecorrF         = gmx_mm256_pmecorrF_ps(zeta2);
 790             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 791             felec            = _mm256_mul_ps(qq00,felec);
 792
 793             /* LENNARD-JONES DISPERSION/REPULSION */
 794
 795             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 796             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 797
 798             fscal            = _mm256_add_ps(felec,fvdw);
 799
 800             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 801
 802             /* Calculate temporary vectorial force */
 803             tx               = _mm256_mul_ps(fscal,dx00);
 804             ty               = _mm256_mul_ps(fscal,dy00);
 805             tz               = _mm256_mul_ps(fscal,dz00);
 806
 807             /* Update vectorial force */
 808             fix0             = _mm256_add_ps(fix0,tx);
 809             fiy0             = _mm256_add_ps(fiy0,ty);
 810             fiz0             = _mm256_add_ps(fiz0,tz);
 811
 812             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 813             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 814             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 815             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 816             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 817             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 818             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 819             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 820             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 821
 822             /* Inner loop uses 64 flops */
 823         }
 824
 825         /* End of innermost loop */
 826
 827         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 828                                                  f+i_coord_offset,fshift+i_shift_offset);
 829
 830         /* Increment number of inner iterations */
 831         inneriter                  += j_index_end - j_index_start;
 832
 833         /* Outer loop uses 7 flops */
 834     }
 835
 836     /* Increment number of outer iterations */
 837     outeriter        += nri;
 838
 839     /* Update outer/inner flops */
 840
 841     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*64);
 842 }